CSFwinformer: Cross-Space-Frequency Window Transformer for Mirror Detection

This repo is the official implementation of CSFwinformer: Cross-Space-Frequency Window Transformer for Mirror Detection (IEEE TIP 2024).

Installation

conda create -n md python=3.7 -y
conda activate md

conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=11.0 -c pytorch
pip install mmcv-full==1.4.0 -f https://download.openmmlab.com/mmcv/dist/cu110/torch1.7.1/index.html

cd CSFwinformer

pip install -e .
pip install -r requirements/optional.txt

mkdir data

Data Preparation

"MSD"

"PMD"

"RGBD-Mirror"

You can download zip files for corresponding three datasets from "here"

Train

python tools/train.py configs/mirror/pmd_mirror_swin_small.py

Test

python ./tools/test.py configs/mirror/pmd_mirror_swin_small.py work_dirs/pmd_mirror_swin_small/your_weight --show-dir ./results/pmd --eval mIoU

Results and Models

Dataset	Backbone	IoU↑	Acc↑	$F_β$↑	MAE↓	BER↓
PMD	swin_s	69.84	77.28	0.849	0.024	11.91
PMD	swin_b	70.05	78.27	0.838	0.024	11.41
MSD	swin_s	82.13	88.72	0.895	0.046	7.15
MSD	swin_b	82.08	88.92	0.896	0.045	7.14
RGBD-Mirror	swin_b	78.66	84.64	0.900	0.031	8.57

You can find all weights from "here"